CT Scan Reconstruction with Neural Fields
Reconstructing a 3D chest CT volume from 50 X-ray projections with learned multi-resolution embeddings.
Python · TensorFlow · NumPy
A COSC440 research project that treats cone-beam CT reconstruction as a differentiable rendering problem, in the spirit of NeRF. Rays are cast from the X-ray source to every detector pixel, a model predicts attenuation at points along each ray, and the integrated values are compared against the measured projections. A chest volume is reconstructed from just 50 projections using a stack of multi-resolution embedding tables feeding a small MLP. I tested two changes to the embedding. Replacing or augmenting the fine tables with neural embeddings sped up early convergence but left final quality within noise. Changing the grid lookup to trilinear interpolation raised SSIM from 0.68 to 0.829 ± 0.025 (PSNR 26.47 dB), and an Instant-NGP style multi-resolution hash grid kept most of that gain (SSIM 0.772) with 14× fewer table parameters. Together these showed that the nearest-cell lookup, not the network, was the main limit on quality. All results are the mean and standard deviation over three seeds, written up as a full report.
